Evaluation of Interpretable Deep Learning Approaches for Neuroimaging in Dementia using Stroke Classification as a Ground Truth for Brain Pathology

痴呆症和阿尔茨海默病是2018年导致死亡的主要原因，在英国的患病率不断上升，预计到2025年将上升到100万。深度学习模型在帮助早期诊断和预后方面显示出巨大的希望。然而，它们的复杂性和缺乏透明度阻碍了它们进入临床实践的可能性。在高风险的临床常规中，这尤其是个问题，误诊的后果可能危及生命。在这些领域，可解释的方法可以提供一种评估和可视化网络行为的方法，以更好地理解为什么会获得特定的结果。然而，由于缺乏验证和缺乏基本真理，最近的方法受到限制。通过利用快速发展的可解释性机器学习领域的最新技术，该项目旨在比较参考中风等“基本事实”神经系统疾病的替代可解释性方法，然后将其应用于神经退行性疾病和痴呆症。目的和目的本研究旨在以脑卒中分类作为脑病理学的基本事实，批判性地验证几种解释痴呆诊断和预后方法的有效性。首先，我们的目标是建立一个网络，以近乎完美的精度将中风患者与对照组区分开来，以确保注意图是有用的和信息丰富的。这将使我们能够探索这些模型在站点和/或数据质量、体素维度和数据规范化方面的通用性。我们将考虑的其他技术因素包括病变大小或萎缩的影响，以及临床特征对模型性能和可解释性的影响。在使用中风的基本事实案例来评估可解释性方法之后，最佳方法将在博士阶段应用于痴呆症。从这些模型得到的注意力图将由具有痴呆症专业知识的放射科医生样本（来自伦敦大学学院和合作站点）进行定性和半定量评估，以供临床使用。研究方法的新颖性可解释医学成像是“可解释人工智能”的一个子领域，是转化研究的一个重要课题。计划中的痴呆症神经图像可解释性研究将是新颖的，并提供队列和患者水平的注意力图，这对于提供精准治疗痴呆症至关重要。据我们所知，使用笔画作为基础真理的可解释性方法的比较评估以前没有报道过。这项研究包括与放射科医生密切合作，提供临床专业知识和临床实用数据，使其成为一个真正的跨学科转化项目。与EPSRC的战略和研究领域保持一致我们的研究重点与EPSRC的战略保持一致，即实现更早、更有效的诊断(1)，并使用机器学习整合来自临床数据和图像的附加信息(3)。此外，与临床医生合作符合与相关利益攸关方密切接触的重点(7)。这项工作属于人工智能技术和图像与视觉计算研究领域。该项目涉及与伦敦大学学院痴呆症研究中心（DRC）和伦敦大学学院医院（UCLH）的NIHR生物医学研究中心的合作。

Brief description of the context of the research including potential impactDementia and Alzheimer's disease were the leading causes of death in 2018, with a growing prevalence in the UK that is expected to rise to 1 million by 2025. Deep learning models have shown great promise in aiding early diagnosis and prognosis. However, their complexity and lack of transparency hinders the likelihood of their adoption into clinical practice. This is particularly a problem in clinical routines where stakes are high, and the consequences of misdiagnosis can be life-threatening. In these areas interpretable approaches can provide a way of assessing, and often visualising, the behaviour of the network to better understand why a particular outcome is obtained. However, recent approaches have been limited due to lack of validation and the absence of a ground truth.By taking advantages of the latest technologies from the rapidly developing field of interpretable ML, this project aims to compare alternative interpretability methods in reference to 'ground truth' neurological conditions like stroke before applying them to neurodegenerative disease and dementia.Aims and ObjectivesThis research aims to critically validate the effectiveness of several interpretation methods for dementia diagnosis and prognosis using stroke classification as a ground truth for brain pathology. To begin with, we will aim to build a network that can classify stroke patients from controls with near-perfect accuracy in order to ensure that the attention maps are useful and informative. This will allow us to explore the generalisability of these models to site and/or data quality, voxel dimension and data normalisation. Other technical considerations we will consider include the effect of lesion size or atrophy, and clinical characteristics on model performance and interpretability. Having used the ground-truth case of stroke to assess interpretability approaches, the optimal approaches will then be applied to dementia duringthe PhD phase. The resulting attention maps from these models will be assessed qualitatively and semi-quantitatively for clinical utility by a sample of radiologists (from UCL and collaborating sites) with expertise in dementia.Novelty of Research MethodologyInterpretable medical imaging, a subfield of 'explainable AI', is an important topic for translational research. The planned research on neuroimage interpretability in dementia will be novel and provide both cohort and patient level attention maps, essential for delivering precision medicine for dementia. To our knowledge, comparative assessment of interpretability methods using stroke as a ground truth has not previously been reported. This research includes working closely with radiologists, to provide clinical expertise and data on clinical utility, making it a truly interdisciplinary translational project.Alignment to EPSRC's strategies and research areasOur research focus aligns with EPSRC's strategy for enabling earlier and more effective diagnosis (1) and integration of additional information from clinical data and images using machine learning (3). In addition, collaborating with clinicians aligns with the focus in strong engagement with relevant stakeholders (7). This work falls within the artificial intelligence technologies and image and vision computing research areas.Any companies or collaborators involvedThe project involves collaboration with the UCL Dementia Research Centre (DRC) and the NIHR Biomedical Research Centre at University College London Hospital (UCLH).